krthkyadtya/Qwen3.8-27B-heretic

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

krthkyadtya/Qwen3.8-27B-heretic is a 27 billion parameter Qwen3.8 model, derived from Qwen/Qwen3.8-27B-FP8, that has undergone "abliteration" to remove refusal behaviors. This model retains the original Qwen3_5ForConditionalGeneration architecture, including a vision tower, and is provided in bf16 format for broad compatibility. It is specifically designed for use cases requiring a large language model with significantly reduced refusal tendencies while largely preserving general capabilities.

Loading preview...

Overview

krthkyadtya/Qwen3.8-27B-heretic is a 27 billion parameter model based on the Qwen3.8 architecture, specifically a modified version of Qwen/Qwen3.8-27B-FP8. This model has been processed using the "Heretic" tool for automatic, directional refusal ablation, effectively "decensoring" it. The original FP8 checkpoint was first dequantized to bf16, and the resulting weights are provided in bf16 format, ensuring compatibility with standard transformers on Ampere/Hopper+ GPUs.

Key Capabilities and Features

  • Refusal Ablation: Successfully reduces refusal behaviors from 30/30 to 0/30 on held-out prompts, as measured by keyword markers.
  • Capability Preservation: Maintains general model capabilities with a low KL divergence of 0.0825 from the original model, well below the "real damage" threshold of ~0.5.
  • Architecture: Preserves the Qwen3_5ForConditionalGeneration architecture, including its hybrid linear + full attention and vision tower components.
  • Compatibility: Provided in bf16 safetensors format, allowing for straightforward loading with transformers without special quantization configurations.
  • Chat Modes: Supports both thinking and non-thinking chat modes.

Use Cases

  • Applications requiring unrestricted responses: Ideal for scenarios where the base model's refusal mechanisms are undesirable.
  • Research into model safety and alignment: Can be used to study the impact of refusal ablation on model behavior and performance.
  • Creative content generation: Suitable for tasks that benefit from a model less prone to self-censorship.

Finetunes
1 models